A face as a point in an eigenvector space
In January 1991 Matthew Turk and Alex Pentland described a system that locates a head in the frame and recognises the person in near real time. A face is projected into the space of the training set's eigenvectors — the "eigenfaces" — and recognition reduces to comparing a handful of coefficients rather than recovering the three-dimensional shape of a nose or an eye.
Why it matters
Face recognition stopped requiring the measurement of features. A face became a vector in a small space and recognition a distance within it; the field worked on that approach until convolutional networks arrived, and it is exactly this class of result that independent evaluation was later built to test.
The publisher's abstract describes the approach this way: the recognition problem is treated as intrinsically two-dimensional, taking advantage of faces normally being upright and so describable by a small set of characteristic views; the system can learn and later recognise new faces in an unsupervised manner. The record states outright what was not checked at first hand. The full text is paywalled and every copy found is an image scan with no text layer, so the figures that follow were read not from the paper but from a Johns Hopkins course handout that cites it by volume and page: 96 percent correct averaged over light variation, 85 percent over orientation variation and 64 percent over size variation, on 16 subjects. That is why the record carries medium confidence. What sits beside it: by NIST's account, before the FERET database existed, face recognition papers usually reported over 95 percent correct on databases of fewer than 50 individuals. Results of exactly that kind are what FERET was built to measure independently.